Every single thing Claude writes now has an invisible watermark baked in. It travels with the text wherever it goes.
TechCrunch reported on Anthropic's new watermarking policy, and the driver is regulatory.
The Trigger
The EU AI Act's Transparency Code, which took effect on August 2, requires AI companies to mark AI-generated or edited content in a way other systems can identify. Non-compliance carries real teeth: fines of up to €15 million or 3% of a company's total global annual turnover, whichever is higher.
How It Actually Works
"Because the watermark is part of the text, it will travel with the text when it's copied and pasted elsewhere, and may persist through some editing. Watermarking will be applied at the model level, which means it will be present no matter which Claude product or surface the text comes from," Anthropic's support page reads.
Under the hood, Anthropic is using an approach called SynthID Text, a technique originally developed by Google DeepMind. Rather than adding any visible marker, it works by nudging Claude's genuinely low-stakes word choices during generation -- picking "overcast" over "grey," for instance -- into a statistical pattern that's invisible to a human reader but detectable by anyone holding the right key. Anthropic says the technique has no impact on output quality: to a reader, watermarked text is "indistinguishable from an unwatermarked one." The company also says it plans to release a separate watermark detection API so third parties -- schools, employers, other platforms -- can check text for the pattern themselves.
The Open Question
It's not clear how much editing users need to do to remove the watermark. Anthropic hasn't clarified exactly where the line sits, though the company has said light editing "probably won't remove the watermark completely," while a full rewrite would -- at which point, Anthropic notes, it's fair to ask whether the result still counts as "AI-generated" at all. Text a person heavily edits and mostly rewrote themselves would likely carry little to no detectable watermark, since the statistical pattern only rides on the words Claude itself chose.
How People Are Reacting
The policy has already sparked real pushback from users worried about being caught by it in low-stakes situations. One commenter framed the concern around students: "The student who used Claude to reorganize a paragraph... Those guys come out of the process with a digital tattoo." Others pushed back on the fairness of it entirely, with one arguing: "Having an AI that watermarks your work is terrifyingly ironic given how many frontier models came by their training data." Defenders of the policy counter that the point isn't to punish light assistance, but to make it possible to "detect AI generated outputs because of the risks AI generated outputs can cause" -- misinformation, impersonation, and academic dishonesty at scale among them.
Industry-Wide
Anthropic isn't acting alone here. Apart from Anthropic, companies including Black Forest Labs, Google, Meta, Microsoft, OpenAI, and Synthesia have all committed to adhering to the EU's transparency code. For image and file-based content specifically, several of these companies -- Anthropic included -- are using the C2PA open standard, which embeds signed provenance information directly into a file rather than relying on a statistical text pattern.
Where It Applies
Watermarking now applies across Claude's platform API, Claude, Claude Code, Claude Cowork, and Claude Tag -- and Anthropic says it will extend support to older models too, and that the markings apply worldwide, not just to users in the EU.